| Ref.No: | 61517200 |
| Start date: | 17.07.2024 |
| End date: | 16.12.2026 |
| Approval date: | 10.07.2024 |
| Department: | ELECTRICAL & COMPUTER ENGINEERING |
| Sector: | COMPUTER SCIENCE |
| Financier: | 5η ΠΡΟΚΗΡΥΞΗ ΕΛΙΔΕΚ Υ.Δ., ELIDEK |
| Budget: | 26.100,00 € |
| Public key: | 90Ο646ΨΖΣ4-1ΝΔ |
| Scientific Responsible: | Prof. STAMOU |
| Email: | gstam@cs.ntua.gr |
| Description: | THE AIM OF THIS THESIS IS TO PROVIDE COMPREHENSIVE EXPLANATIONS OF AI SYSTEMS, FOCUSING ON UTILIZING GRAPH STRUCTURES AS INPUT AND EMPLOYING GRAPH NEURAL NETWORKS (GNNS). THE THESIS ADDRESSES THE CHALLENGE OF EXPLAINING "BLACK BOX" SYSTEMS BY OFFERING UNIVERSAL, COMPLETE, AND INTERPRETABLE APPROACHES. IT BEGINS WITH EXPLAINING IMAGE CLASSIFIERS USING COUNTERFACTUAL EXPLANATIONS WITHIN THE CONTEXT OF COMPUTER VISION, LEVERAGING SCENE GRAPHS OF IMAGES. A COUNTERFACTUAL FRAMEWORK WILL BE DEVELOPED FOR VISUAL GENOME AND CUB 200 2011 DATASETS. SUBSEQUENTLY, THE FRAMEWORK WILL BE APPLIED TO AUDIO AND TEXTUAL DATA USING THE WEBNLG DATASET. THE THESIS WILL EVALUATE THE EFFECTIVENESS AND QUALITY OF THE EXPLANATIONS, ESTABLISH A FORMAL FRAMEWORK FOR EVALUATING COUNTERFACTUALS, AND EXPLORE WHETHER GNNS PROVIDE INTRINSIC INTERPRETABILITY. THE GOAL IS TO DEVELOP EFFICIENT AND HUMAN-UNDERSTANDABLE TOOLS FOR DOMAINS REQUIRING PRECISE CONTROL OVER PREDICTIONS. |
